基于卷积神经网络识别重力异常体  

The identification of gravity anomaly body based on the convolutional neural network

作　　者：王逸宸 柳林涛[1,2] 许厚泽[1,2] WANG Yi-Chen;LIU Lin-Tao;XU Hou-Ze(Institute of Geodesy and Geophysics,Chinese Academy of Sciences,Wuhan 430077,China;State Key Laboratory of Geodesy and Earth s Dynamics,Chinese Academy of Sciences,Wuhan 430077,China;Collage of Earth and Planetary Sciences,University of Chinese Academy of Sciences,Beijing100049,China)

机构地区：[1]中国科学院测量与地球物理研究所,湖北武汉430077 [2]中国科学院大地测量与地球动力学国家重点实验室,湖北武汉430077 [3]中国科学院大学地球与行星科学学院,北京100049

出　　处：《物探与化探》2020年第2期394-400,共7页Geophysical and Geochemical Exploration

基　　金：国家自然科学基金项目(Y211641064);国家重大科学仪器设备开发专项基金项目“海洋/航空重力仪研制”(20011YQ120045)。

摘　　要：本文将深度学习与重力异常体识别结合,基于近年来在图像识别邻域取得优异效果的卷积神经网络,将重力观测等值线图看作待识别的二维图像,将地下重力异常体的空间参数看作识别输出,从而形成适用于异常体识别的卷积神经网络模型。在训练中,随机生成大量不同参数的三维异常体模型,正演得到其重力观测二维数据,用异常体模型参数标签和重力数据训练卷积神经网络。在模型算例中测试训练好的网络模型,其识别准确性良好。同时,相比于传统神经网络从二维重力测线中识别异常体的埋深,卷积神经网络可从二维的重力数据识别三维异常体的埋深和大小信息。最后,将网络应用于澳大利亚Kauring地区重力观测数据,异常体识别结果与前人研究结果相符。说明卷积神经网络具泛化能力,可用于识别实测重力异常体,结果可靠。This study combines the deep learning with the identification of gravity anomaly body. Based on the CNN(convolutional neural network) which has been gaining its use in the past several years in the field of image identification, the contour image of gravity signal is taken as the unidentified image, while the space parameters of the gravity anomaly body will be identified through CNN. In the training phase, a large number of the 3 D anomaly bodies are generated with random variation of parameters, then the network is fed with parametric labels and the computed gravity contour images. The testing is performed with generated testing models to estimate the performance of the trained model. The trained CNN accuracy shows excellent accuracy in the identifications. Then the CNN model is tested with measured main gravity anomaly data of Kauring area in West Australia, and the identified parameters of the 3 D anomaly body are compared with known results. It is shown that the generalization of CNN can handle identification of the measured gravity data.

关 键 词：深度学习 卷积神经网络 重力异常体识别 参数反演 

分 类 号：P631[天文地球—地质矿产勘探]